XVIII IWRA World Water Congress Beijing China 2023
Sub-theme 3: Building Resilience for Disaster Prevention and Mitigation
Author(s): Miss. Ruonan Hao, Anhui University of Science & Technology
Miss. Ruonan Hao, Anhui University of Science & Technology
Co-author(s): Dr. Zhixu Bai, Wenzhou Univeristy
Keyword(s): Daily streamflow simulating, Machine learning, LSTM, XGBoost, Bayesian optimization
Oral: PDFAbstractRainfall-runoff modelling has been of great importance for flood control and water resource management. However, there are still some problems in the selection of hydrological models to obtain satisfactory simulating performance. With the rapid development of machine learning techniques in hydrology, the performances of three modelling methods under different categories of machine learning, including Support Vector Regression (SVR), eXtreme Gradient Boosting (XGBoost), and Long-Short Term Memory neural network (LSTM), were assessed in Daitou catchment, a mountainous river in Zhejiang Province of China, in terms of daily runoff simulating. The parameters of these models were optimized by the Bayesian optimization method. Besides, simulating performances of the three models with different input scenarios were compared to analyze the impacts of antecedent streamflow and distribution of rainfall. All the performances were assessed by Nash-Sutclife efficiency (NSE), root mean square error (RMSE), and correlation coefficient (CC). The results show that (1) LSTM always obtained a higher accuracy than XGBoost, followed by SVR, and the distribution of rainfall and antecedent streamflow facilitated the improvement of simulating with the best NSE value of 0.75 for LSTM; (2) With the best input scenario, XGBoost (with NSE of 0.58) showed a better performance than LSTM (with NSE of 0.39) during dry seasons when trained with all data including data during wet seasons; (3) the classification of wet and dry seasons improved the simulating performance, especially for LSTM during dry seasons, which indicated distinction of various rainfall-runoff processes can prompt the robust and satisfactory modelling; (4) the different performances in validation and testing periods compared with that in training period may depend on the linear correlation between input and output. The significantly different results obtained by the three machine learning methods revealed their different characteristics when capturing the streamflow fluctuations, providing some insights for the selection of machine learning methods during wet and dry seasons.